From patchwork Tue Mar 17 20:06:51 2015 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Don Brace X-Patchwork-Id: 6034401 Return-Path: X-Original-To: patchwork-linux-scsi@patchwork.kernel.org Delivered-To: patchwork-parsemail@patchwork1.web.kernel.org Received: from mail.kernel.org (mail.kernel.org [198.145.29.136]) by patchwork1.web.kernel.org (Postfix) with ESMTP id 21DE79F2A9 for ; Tue, 17 Mar 2015 20:08:44 +0000 (UTC) Received: from mail.kernel.org (localhost [127.0.0.1]) by mail.kernel.org (Postfix) with ESMTP id D8A6220459 for ; Tue, 17 Mar 2015 20:08:42 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id 5517B20480 for ; Tue, 17 Mar 2015 20:08:40 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932459AbbCQUIg (ORCPT ); Tue, 17 Mar 2015 16:08:36 -0400 Received: from g9t1613g.houston.hp.com ([15.240.0.71]:42597 "EHLO g9t1613g.houston.hp.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932442AbbCQUIb (ORCPT ); Tue, 17 Mar 2015 16:08:31 -0400 Received: from g9t5008.houston.hp.com (g9t5008.houston.hp.com [15.240.92.66]) (using TLSv1.2 with cipher DHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by g9t1613g.houston.hp.com (Postfix) with ESMTPS id B74036423F for ; Tue, 17 Mar 2015 20:08:30 +0000 (UTC) Received: from g4t3433.houston.hp.com (g4t3433.houston.hp.com [16.210.25.219]) by g9t5008.houston.hp.com (Postfix) with ESMTP id 49AD5148; Tue, 17 Mar 2015 20:08:30 +0000 (UTC) Received: from [127.0.1.1] (brunhilda.americas.hpqcorp.net [16.100.201.25]) by g4t3433.houston.hp.com (Postfix) with ESMTP id 26E65B1; Tue, 17 Mar 2015 20:08:30 +0000 (UTC) Subject: [PATCH v3 37/42] hpsa: use block layer tag for command allocation From: Don Brace To: scott.teel@pmcs.com, Kevin.Barnett@pmcs.com, james.bottomley@parallels.com, hch@infradead.org, Justin.Lindley@pmcs.com, brace@pmcs.com Cc: linux-scsi@vger.kernel.org Date: Tue, 17 Mar 2015 15:06:51 -0500 Message-ID: <20150317200651.19856.77450.stgit@brunhilda> In-Reply-To: <20150317200139.19856.87982.stgit@brunhilda> References: <20150317200139.19856.87982.stgit@brunhilda> User-Agent: StGit/0.17.1-dirty MIME-Version: 1.0 Sender: linux-scsi-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-scsi@vger.kernel.org X-Spam-Status: No, score=-6.9 required=5.0 tests=BAYES_00, RCVD_IN_DNSWL_HI, T_RP_MATCHES_RCVD, UNPARSEABLE_RELAY autolearn=unavailable version=3.3.1 X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on mail.kernel.org X-Virus-Scanned: ClamAV using ClamSMTP From: Webb Scales Rework slave allocation: - separate the tagging support setup from the hostdata setup - make the hostdata setup act consistently when the lookup fails - make the hostdata setup act consistently when the device is not added - set up the queue depth consistently across these scenarios - if the block layer mq support is not available, explicitly enable and activate the SCSI layer tcq support (and do this at allocation-time so that the tags will be available for INQUIRY commands) Tweak slave configuration so that devices which are masked are also not attached. Reviewed-by: Scott Teel Reviewed-by: Kevin Barnett Signed-off-by: Webb Scales Signed-off-by: Don Brace --- drivers/scsi/hpsa.c | 153 +++++++++++++++++++++++++++++++++++++++++---------- drivers/scsi/hpsa.h | 1 2 files changed, 123 insertions(+), 31 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html diff --git a/drivers/scsi/hpsa.c b/drivers/scsi/hpsa.c index 34c178c..4e34a62 100644 --- a/drivers/scsi/hpsa.c +++ b/drivers/scsi/hpsa.c @@ -44,6 +44,7 @@ #include #include #include +#include #include #include #include @@ -212,6 +213,9 @@ static int hpsa_compat_ioctl(struct scsi_device *dev, int cmd, static void cmd_free(struct ctlr_info *h, struct CommandList *c); static struct CommandList *cmd_alloc(struct ctlr_info *h); +static void cmd_tagged_free(struct ctlr_info *h, struct CommandList *c); +static struct CommandList *cmd_tagged_alloc(struct ctlr_info *h, + struct scsi_cmnd *scmd); static int fill_cmd(struct CommandList *c, u8 cmd, struct ctlr_info *h, void *buff, size_t size, u16 page_code, unsigned char *scsi3addr, int cmd_type); @@ -2047,11 +2051,17 @@ static void hpsa_cmd_resolve_events(struct ctlr_info *h, } } +static void hpsa_cmd_resolve_and_free(struct ctlr_info *h, + struct CommandList *c) +{ + hpsa_cmd_resolve_events(h, c); + cmd_tagged_free(h, c); +} + static void hpsa_cmd_free_and_done(struct ctlr_info *h, struct CommandList *c, struct scsi_cmnd *cmd) { - hpsa_cmd_resolve_events(h, c); - cmd_free(h, c); + hpsa_cmd_resolve_and_free(h, c); cmd->scsi_done(cmd); } @@ -2072,8 +2082,7 @@ static void hpsa_cmd_abort_and_free(struct ctlr_info *h, struct CommandList *c, hpsa_set_scsi_cmd_aborted(cmd); dev_warn(&h->pdev->dev, "CDB %16phN was aborted with status 0x%x\n", c->Request.CDB, c->err_info->ScsiStatus); - hpsa_cmd_resolve_events(h, c); - cmd_free(h, c); /* FIX-ME: change to cmd_tagged_free(h, c) */ + hpsa_cmd_resolve_and_free(h, c); } static void process_ioaccel2_completion(struct ctlr_info *h, @@ -4535,7 +4544,7 @@ static int hpsa_ciss_submit(struct ctlr_info *h, } if (hpsa_scatter_gather(h, c, cmd) < 0) { /* Fill SG list */ - cmd_free(h, c); + hpsa_cmd_resolve_and_free(h, c); return SCSI_MLQUEUE_HOST_BUSY; } enqueue_cmd_and_start_io(h, c); @@ -4581,6 +4590,8 @@ static inline void hpsa_cmd_partial_init(struct ctlr_info *h, int index, { dma_addr_t cmd_dma_handle = h->cmd_pool_dhandle + index * sizeof(*c); + BUG_ON(c->cmdindex != index); + memset(c->Request.CDB, 0, sizeof(c->Request.CDB)); memset(c->err_info, 0, sizeof(*c->err_info)); c->busaddr = (u32) cmd_dma_handle; @@ -4675,27 +4686,24 @@ static int hpsa_scsi_queue_command(struct Scsi_Host *sh, struct scsi_cmnd *cmd) /* Get the ptr to our adapter structure out of cmd->host. */ h = sdev_to_hba(cmd->device); + + BUG_ON(cmd->request->tag < 0); + dev = cmd->device->hostdata; if (!dev) { cmd->result = DID_NO_CONNECT << 16; cmd->scsi_done(cmd); return 0; } - memcpy(scsi3addr, dev->scsi3addr, sizeof(scsi3addr)); - if (unlikely(lockup_detected(h))) { - cmd->result = DID_NO_CONNECT << 16; - cmd->scsi_done(cmd); - return 0; - } - c = cmd_alloc(h); + memcpy(scsi3addr, dev->scsi3addr, sizeof(scsi3addr)); if (unlikely(lockup_detected(h))) { cmd->result = DID_NO_CONNECT << 16; - cmd_free(h, c); cmd->scsi_done(cmd); return 0; } + c = cmd_tagged_alloc(h, cmd); /* * Call alternate submit routine for I/O accelerated commands. @@ -4708,7 +4716,7 @@ static int hpsa_scsi_queue_command(struct Scsi_Host *sh, struct scsi_cmnd *cmd) if (rc == 0) return 0; if (rc == SCSI_MLQUEUE_HOST_BUSY) { - cmd_free(h, c); + hpsa_cmd_resolve_and_free(h, c); return SCSI_MLQUEUE_HOST_BUSY; } } @@ -4822,15 +4830,23 @@ static int hpsa_register_scsi(struct ctlr_info *h) sh->hostdata[0] = (unsigned long) h; sh->irq = h->intr[h->intr_mode]; sh->unique_id = sh->irq; + error = scsi_init_shared_tag_map(sh, sh->can_queue); + if (error) { + dev_err(&h->pdev->dev, + "%s: scsi_init_shared_tag_map failed for controller %d\n", + __func__, h->ctlr); + goto fail_host_put; + } error = scsi_add_host(sh, &h->pdev->dev); - if (error) + if (error) { + dev_err(&h->pdev->dev, "%s: scsi_add_host failed for controller %d\n", + __func__, h->ctlr); goto fail_host_put; + } scsi_scan_host(sh); return 0; fail_host_put: - dev_err(&h->pdev->dev, "%s: scsi_add_host" - " failed for controller %d\n", __func__, h->ctlr); scsi_host_put(sh); return error; fail: @@ -4840,6 +4856,23 @@ static int hpsa_register_scsi(struct ctlr_info *h) } /* + * The block layer has already gone to the trouble of picking out a unique, + * small-integer tag for this request. We use an offset from that value as + * an index to select our command block. (The offset allows us to reserve the + * low-numbered entries for our own uses.) + */ +static int hpsa_get_cmd_index(struct scsi_cmnd *scmd) +{ + int idx = scmd->request->tag; + + if (idx < 0) + return idx; + + /* Offset to leave space for internal cmds. */ + return idx += HPSA_NRESERVED_CMDS; +} + +/* * Send a TEST_UNIT_READY command to the specified LUN using the specified * reply queue; returns zero if the unit is ready, and non-zero otherwise. */ @@ -4979,18 +5012,18 @@ static int hpsa_eh_device_reset_handler(struct scsi_cmnd *scsicmd) /* if controller locked up, we can guarantee command won't complete */ if (lockup_detected(h)) { dev_warn(&h->pdev->dev, - "scsi %d:%d:%d:%d RESET FAILED, lockup detected\n", - h->scsi_host->host_no, dev->bus, dev->target, - dev->lun); + "scsi %d:%d:%d:%u cmd %d RESET FAILED, lockup detected\n", + h->scsi_host->host_no, dev->bus, dev->target, dev->lun, + hpsa_get_cmd_index(scsicmd)); return FAILED; } /* this reset request might be the result of a lockup; check */ if (detect_controller_lockup(h)) { dev_warn(&h->pdev->dev, - "scsi %d:%d:%d:%d RESET FAILED, new lockup detected\n", + "scsi %d:%d:%d:%u cmd %d RESET FAILED, new lockup detected\n", h->scsi_host->host_no, dev->bus, dev->target, - dev->lun); + dev->lun, hpsa_get_cmd_index(scsicmd)); return FAILED; } @@ -5442,6 +5475,59 @@ static int hpsa_eh_abort_handler(struct scsi_cmnd *sc) } /* + * For operations with an associated SCSI command, a command block is allocated + * at init, and managed by cmd_tagged_alloc() and cmd_tagged_free() using the + * block request tag as an index into a table of entries. cmd_tagged_free() is + * the complement, although cmd_free() may be called instead. + */ +static struct CommandList *cmd_tagged_alloc(struct ctlr_info *h, + struct scsi_cmnd *scmd) +{ + int idx = hpsa_get_cmd_index(scmd); + struct CommandList *c = h->cmd_pool + idx; + int refcount = 0; + + if (idx < HPSA_NRESERVED_CMDS || idx >= h->nr_cmds) { + dev_err(&h->pdev->dev, "Bad block tag: %d not in [%d..%d]\n", + idx, HPSA_NRESERVED_CMDS, h->nr_cmds - 1); + /* The index value comes from the block layer, so if it's out of + * bounds, it's probably not our bug. + */ + BUG(); + } + + refcount = atomic_inc_return(&c->refcount); + if (unlikely(!hpsa_is_cmd_idle(c))) { + /* + * We expect that the SCSI layer will hand us a unique tag + * value. Thus, there should never be a collision here between + * two requests...because if the selected command isn't idle + * then someone is going to be very disappointed. + */ + dev_err(&h->pdev->dev, + "tag collision (tag=%d) in cmd_tagged_alloc().\n", + idx); + if (c->scsi_cmd != NULL) + scsi_print_command(c->scsi_cmd); + scsi_print_command(scmd); + } + + hpsa_cmd_partial_init(h, idx, c); + return c; +} + +static void cmd_tagged_free(struct ctlr_info *h, struct CommandList *c) +{ + /* + * Release our reference to the block. We don't need to do anything + * else to free it, because it is accessed by index. (There's no point + * in checking the result of the decrement, since we cannot guarantee + * that there isn't a concurrent abort which is also accessing it.) + */ + (void)atomic_dec(&c->refcount); +} + +/* * For operations that cannot sleep, a command block is allocated at init, * and managed by cmd_alloc() and cmd_free() using a simple bitmap to track * which ones are free or in use. Lock must be held when calling this. @@ -5454,7 +5540,6 @@ static struct CommandList *cmd_alloc(struct ctlr_info *h) { struct CommandList *c; int refcount, i; - unsigned long offset; /* * There is some *extremely* small but non-zero chance that that @@ -5466,31 +5551,39 @@ static struct CommandList *cmd_alloc(struct ctlr_info *h) * very unlucky thread might be starved anyway, never able to * beat the other threads. In reality, this happens so * infrequently as to be indistinguishable from never. + * + * Note that we start allocating commands before the SCSI host structure + * is initialized. Since the search starts at bit zero, this + * all works, since we have at least one command structure available; + * however, it means that the structures with the low indexes have to be + * reserved for driver-initiated requests, while requests from the block + * layer will use the higher indexes. */ - offset = h->last_allocation; /* benignly racy */ for (;;) { - i = find_next_zero_bit(h->cmd_pool_bits, h->nr_cmds, offset); - if (unlikely(i == h->nr_cmds)) { - offset = 0; + i = find_first_zero_bit(h->cmd_pool_bits, HPSA_NRESERVED_CMDS); + if (unlikely(i >= HPSA_NRESERVED_CMDS)) continue; - } c = h->cmd_pool + i; refcount = atomic_inc_return(&c->refcount); if (unlikely(refcount > 1)) { cmd_free(h, c); /* already in use */ - offset = (i + 1) % h->nr_cmds; continue; } set_bit(i & (BITS_PER_LONG - 1), h->cmd_pool_bits + (i / BITS_PER_LONG)); break; /* it's ours now. */ } - h->last_allocation = i; /* benignly racy */ hpsa_cmd_partial_init(h, i, c); return c; } +/* + * This is the complementary operation to cmd_alloc(). Note, however, in some + * corner cases it may also be used to free blocks allocated by + * cmd_tagged_alloc() in which case the ref-count decrement does the trick and + * the clear-bit is harmless. + */ static void cmd_free(struct ctlr_info *h, struct CommandList *c) { if (atomic_dec_and_test(&c->refcount)) { diff --git a/drivers/scsi/hpsa.h b/drivers/scsi/hpsa.h index 3ec8934..2536b67 100644 --- a/drivers/scsi/hpsa.h +++ b/drivers/scsi/hpsa.h @@ -141,7 +141,6 @@ struct ctlr_info { struct CfgTable __iomem *cfgtable; int interrupts_enabled; int max_commands; - int last_allocation; atomic_t commands_outstanding; # define PERF_MODE_INT 0 # define DOORBELL_INT 1